AI researcher and builder

Parker Fawcett

I research how AI systems fail, and I build with them.

I write papers on how to test AI systems, build products and agent tools, and contribute to open source. I report failures next to passes. By day I build AI tooling at CHG Healthcare, and I run Skora, a college admissions platform.

Read the researchRead the writing

Portrait of Parker Fawcett

Research

How AI gets checked.

Three papers so far, on coding-agent verification and on machine-learning evaluation in finance. Each one reports what failed.

Read the research →

Build

Products and agents.

A college admissions platform, a collectibles marketplace, an AI receptionist, a business stress-test tool and a published Roblox game.

See the projects →

Open source

Tools in public.

Merged fixes to the OpenClaw agent framework, small agent tools such as custody and Skill-Drift, talks, and a youth coding competition I ran.

See the code →

Selected result

A passing test suite is not a correct rebuild.

A coding agent can rebuild an app until every test goes green. When one pipeline writes both the spec and the tests, they can share the same blind spot, so green says less than it appears to. My work measures that gap and proposes ways to audit it.

See the papers →

An HTTP 200 stub, scored two ways
  • Visible tests passed20/20
  • Held-out test files passed12/12
  • Behavioral checks failed0/77

Reproduce the baseline ↗

A stock-picking model, corrected for leakage
  • Test AUC held steady0.553 → 0.551
  • Selected stocks in common changed38.9%

Earlier version on SSRN ↗

Latest writing

All writing →

Also in progress: Experience Without Permission, a book on how AI changes the way beginners become experts.

Preprint on arXiv Software engineering, agentic systems, 2026

Rebuild Dossier: Mechanically-Enforced Specs for Agentic App Rebuilds

An open-source MCP server that extracts contracts and tests from an existing app so a fresh coding agent can rebuild it from a spec. The paper reports the runs that worked and the ones that failed.

Read on arXiv ↗Code and evidence ↗Archived software DOI ↗

Under review at ICSE 2027 SEIP Software engineering, agentic systems, 2026

Shared Blind Spots in Co-Generated Specifications and Acceptance Tests

When one pipeline writes both the spec and the acceptance tests, they share a blind spot. A do-nothing stub passed every generated test across 30 frozen rebuilds. The paper proposes a three-step audit to run before trusting a green rebuild.

Code and evidence ↗

Under review at the Journal of Financial Data Science Quantitative finance, machine learning, 2026

When AUC Survives but Portfolios Do Not

A model’s accuracy score barely moved (0.553 to 0.551), yet 61.1% of the stocks it picked changed. An audit of how a stable metric can sit on top of unstable decisions.

Earlier version on SSRN ↗Code and evidence ↗

Active research Open-source research tool, 2026

Rebuild Dossier

Records what an app does before an AI agent rebuilds it, so the agent has a spec to work from and a test suite you can audit. The papers report where that process catches problems and where a green run still hides them.

Source ↗Paper ↗Research page →

Skora admissions platform screenshot

Early stage, in counselor pilots College admissions platform, 2025 to present

Skora

Students see their estimated chances at each college, get matched to schools and get feedback on essays. Counselors manage their students and offer Skora under their own brand. College facts come from College Scorecard and IPEDS data. Essay feedback is drafted by AI and reviewed by a counselor. I am the CEO.

skoraadmit.com ↗

CatchAndTrade marketplace screenshot

Live Collectibles marketplace, 2025 to present

CatchAndTrade

A marketplace and portfolio tracker for Pokémon card collectors. Search a card database, track what a collection is worth, and trade or sell with other collectors. Card scanning runs on the collector’s own device. It is also the app behind the 30 frozen rebuilds in the Rebuild Dossier study.

catchandtrade.com ↗

MyNexusAI receptionist screenshot

Prototype AI receptionist, 2025 to present

MyNexusAI

Answers a small business’s calls and texts using the business’s own documents. It started when an HVAC chatbot I was testing invented a coupon that appeared nowhere in its source material. That turned my attention to grounding and AI reliability.

mynexusai.org ↗

Lightning talk: how easier AI-enabled building changes the development of professional judgment

University of Utah Student-Led AI Symposium. Gould Auditorium, Salt Lake City, session on AI and the Future of Work.

Five minutes on why building has become easier, why judgment still takes real-world feedback, and what that means for how beginners learn.

How we test whether an AI agent rebuilt an app correctly

Vibe Code Club, Utah Valley Business Resource Center. Main research presentation, 15 minutes.

The Rebuild Dossier results on agent rebuild testing, including the stub that passed every generated test.

Get in touch

Email is the fastest way to reach me. I welcome replication attempts, technical feedback and invitations to speak or write.

Parkerscottfawcett@gmail.comPress kit